DViTPolyp: Small Polyp Detection in Colonoscopy Videos based on Deep Vision Transformer

Le résumé fourni par la source

Early detection of colorectal polyps and removal play a crucial role in the prevention of colon cancer. Computer-aided detection (CAD) methods can be employed to help endoscopists reduce the polyp miss-rates. Numerous automatic polyp detection methods have been explored. However, it is still challenging to automatically detect colon polyps in small size in colonoscopy videos. This paper presents a method termed DViTPolyp for detecting small colorectal polyp in colonoscopy videos. The proposed method leverages the power of the deep vision transformer as the detection head in the network to aggregate rich global features, and uses the re-attention mechanism to regenerate attention maps to capture the features from the backbone model and DViT block. Also, we propose to use a power regularized Intersection Over Union (IoU) as the loss function to further enhance the accuracy of detection. Experiments are conducted on a mix of several public polyp datasets. Comparative results show that our proposed method has a better performance than other methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.